What are the signs of AI writing?
Most people look for the signs of AI writing in the style. They watch for a stock vocabulary, inflated statements of importance, “not just X, but Y” constructions, opinions attributed to unnamed experts, a formulaic outline, and leftovers from the chat window. None of them proves a text was written by a model, because people write this way too.
A second group gets far less attention, and it sits in the evidence. Here the signs are a quotation the source does not contain, a number that measures something else, a figure from a replaced edition, a second-hand source, and a reference that leads nowhere. Readers cannot see them without opening the sources. They are also the ones that turn out to be wrong in a way somebody can prove.
Both groups are covered below, eleven signs in all, with a check for each of the hidden five. We wrote the list for people who approve business content before it goes live, and it ends with a routine that takes about ten minutes per draft.
Six signs you can see
Wikipedia keeps the best public catalog of these, a project page called Wikipedia:Signs of AI writing. Its editors describe it as “a list of writing and formatting conventions typical of AI chatbots” with examples taken from the site itself. Our six follow its main groups, and the phrases in quotation marks are ones the page lists as words to watch.
1. A stock vocabulary
Certain words turn up far more often in model output than in comparable human text. Wikipedia’s page lists “Additionally” at the start of a sentence, “pivotal”, “robust” and “showcase” among them. You can measure the effect at scale. One study of more than 15 million biomedical abstracts counted the words that became suddenly more frequent and concluded that “at least 13.5% of 2024 abstracts were processed with LLMs.”
2. Inflated significance
Ordinary subjects get described as historic. Examples on the page include “is a testament”, “evolving landscape” and “indelible mark”. What gives it away is the gap between the size of the claim and the size of the thing, as when a routine release note is said to change an industry.
3. Negative parallelisms and the rule of three
Two sentence shapes repeat. One denies something nobody said in order to assert something else, as in “Not just X, but also Y”. Its companion groups everything in threes, whether or not there are three things. Models often use the triple, the page notes, “to make superficial analyses appear more comprehensive.”
4. Opinions from nobody in particular
Claims are attributed to an authority that is never named. Wikipedia’s examples are “Industry reports”, “Experts argue” and “Observers have cited”. Such a sentence sounds sourced and is not. We place this sign on the border between the two groups, because the repair is the one the hidden signs need. Name the source or cut the claim.
A sentence that says experts agree, and names none, sounds sourced and is not.
5. A formulaic outline
Whatever the subject, the piece follows one template. Its introduction restates the title, its sections run to equal length, and its closing section begins “Despite these challenges” before ending on a vaguely hopeful note. At sentence level the same evenness shows as similar lengths and paragraphs that open the same way. Researchers who studied how people spot AI text found that after vocabulary, the most common giveaways were formulaic sentence and document structures (Russell, Karpinska and Iyyer, 2025).
6. Leftovers from the chat window
Sometimes the draft still talks to the person who asked for it (“Here is your Wikipedia article on…”). It may carry a disclaimer about a training cutoff, or an unfilled placeholder such as “[Birth Date]”. Only these visible signs are treated by Wikipedia as close to conclusive. Its speedy-deletion rule for unambiguously LLM-generated pages names “Communication intended for the user” as one of three grounds. Another of the three is about references, and it belongs to the hidden group below.
Why the visible signs prove so little
Before you act on a visible sign, consider how people and tools perform at reading them.
People are poor judges. In the Russell study, annotators read 300 non-fiction articles and labeled each as human-written or AI-generated. Those who did not use AI tools for writing detected AI text “at a similar rate to random chance.” Five annotators who use such tools heavily did far better, and their majority vote “misclassifies only 1 of 300 articles.” So the skill exists, but most reviewers do not have it, and confidence is no guide to who does.
Detectors make the same mistake in the other direction. Stanford researchers ran seven detectors over 91 essays written by non-native English speakers for the TOEFL exam, and the tools “misclassified over half of the TOEFL essays” as AI-generated, an “average false positive rate: 61.22%” (Liang and colleagues, 2023). OpenAI withdrew its own classifier within six months of launch. Its announcement had reported that the tool caught 26% of AI-written text while flagging human text 9% of the time, and now carries the note that it “is no longer available due to its low rate of accuracy.” We cover the mechanics in how AI detectors work and what a score can and cannot settle in AI detector vs citation check.
Signs also expire. Once a tell becomes famous, writers and model makers remove it. One analysis of arXiv abstracts found that words flagged as ChatGPT favorites, “delve” among them, dropped in frequency “starting soon after they were pointed out in early 2024.” Wikipedia now files the em dash, long one of the best-known tells, under “Historical indicators”, noting that “some AI companies have attempted to make their newer chatbots suppress their use.”
About its own list the page says, in bold: “Please do not merely treat these signs as the problems to be fixed; that could just make detection harder.” It describes the patterns as “only potential signs of a problem, not the problem itself.” For that problem, read the next five.
A draft can lose every visible sign of AI writing and still carry a figure its source never published.
Five signs you can’t see
None of these five shows in the prose, and a draft can read well and still have all of them. Google explains why they are common in AI-assisted drafts in its guidance on generative AI content: “generative models don’t retrieve facts, but predict a likely sequence of words based on their training data” (Google Search Central). Models produce the kind of sentence that usually follows, and a citation is one more sentence.
We know these signs from our own drafts. Nine articles were published on this blog before this one, each drafted with AI assistance and then checked claim by claim. Of 314 claims opened on their sources, 33 (10.5%) had to be corrected, in seven of the nine articles, and 16 of those corrections came after publication. Every example below is one of the 33.
7. A quotation or figure the source does not contain
Here the source is real and the link opens, but the words or the number are not on the page. Measured rates are high even when the model can search. One 2025 study in Nature Communications tested seven models on 800 medical questions and found that “between 50% and 90% of LLM responses are not fully supported, and sometimes contradicted, by the sources they cite.” With live web search the result was better and still poor, since “approximately 30% of individual statements are unsupported” (Wu and colleagues, 2025).
Ours was milder and typical. One draft of our fact-checking article quoted OpenAI inside quotation marks but had silently dropped a clause from the middle of the sentence. Elsewhere in that draft, social media users were said to have spotted an error “immediately” where the cited report said they “quickly pointed out” the mistake. Neither error changes the story. But quotation marks tell the reader these are the source’s exact words, and in both cases they were not.
Check: open the source and search the page for five consecutive words of the quotation, or for the number. If the search finds nothing, the claim goes or gets rewritten to what the page does say.
8. A number that is on the page but measures something else
Sometimes the figure is real and the draft attaches it to the wrong thing, such as a different population or a different denominator. Errors of this kind survive a quick check, because the number is there when you search for it.
One draft of our article on content ROI gave a study’s 42% as a share of all 7,683 pages the study had dated. On the page it is a share of the 4,124 pages where both the publish date and the update date could be read. Both versions cite one source and one number, yet they make different claims.
Check: read the sentence around the number on the source page and write down what was counted, out of what, and when. If your sentence and that note disagree, yours is wrong.
9. A figure from an edition that has been replaced
Annual surveys and reports are reissued, and a model’s knowledge stops at its training date. So a draft cites last year’s edition with full confidence, and nothing in the prose shows it.
When we drafted our article on agency margins, a blogging-time figure came from a 2025 survey of 808 marketers. By the time we checked, the same URL carried the 2026 edition, with 1,042 respondents and a different number. Another rate in that draft was presented as coming from a report published in November 2025, when the survey behind it described respondents as of November 2021.
Facts go stale after the check too. Two of our published articles were out of date within 18 days of their last review, because two AI vendors changed what their products do. That is why the date of the check belongs on the page.
Two of our published articles were out of date within 18 days of their last review.
Check: find the fieldwork or publication date on the source page, then search for the report’s name with the current year. For anything about a product, read the vendor’s own page on the day you publish.
10. A second-hand source standing in for the study
Often a draft cites an article about a study where it could cite the study. Each retelling can shift a number or its meaning, and a model tends to cite whatever page states the claim most quotably.
People have always done this, which is why models learned it. One meta-analysis of 28 studies of medical journal articles estimated that 25.4% of quotations contained an error against the cited source. According to a much-quoted 2002 estimate, modeled from the way misprints spread through reference lists, “only about 20% of citers read the original.”
Check: follow the links until you reach the page that first published the number. Cite that page. If you cannot reach it, say who reported the figure instead of presenting it as established.
11. A reference that leads nowhere
Models can produce a citation with plausible authors, title and journal for a paper that does not exist, or a link that was never live. Such a reference stays hidden only until someone clicks, and most readers never do. When researchers asked ChatGPT for literature reviews in a 2023 test, “55% of the GPT-3.5 citations but just 18% of the GPT-4 citations are fabricated,” and many of the real ones carried errors. Newer models with search do better. Even so, Wikipedia still lists the pattern under Citations and says that when several links in a new article are broken and no archive has them, “this is a strong sign of an AI-generated page.” Its deletion rule names “Non-existent or nonsensical references” as a second ground.
Check: click every link. For a paper, resolve the DOI and confirm that the title and authors on the landing page match the citation.
All eleven signs at a glance
| # | Sign | Who can see it | What it proves | How to check |
|---|---|---|---|---|
| 1 | Stock vocabulary | Any reader | Nothing on its own | Read aloud, then rewrite |
| 2 | Inflated significance | Any reader | Nothing on its own | Ask who says so |
| 3 | Negative parallelisms, rule of three | Any reader | Nothing on its own | Rewrite the sentence |
| 4 | Unnamed authorities | Any reader | A missing source | Name the source or cut |
| 5 | Formulaic outline | Any reader | Nothing on its own | Restructure around the reader’s question |
| 6 | Chat leftovers | Any reader | Unreviewed model output | Delete, then review the whole draft |
| 7 | Quotation or figure not in the source | Only a checker | A false statement | Search the source for the words |
| 8 | Number that measures something else | Only a checker | A false statement | Read what was counted, out of what |
| 9 | Figure from a replaced edition | Only a checker | An outdated statement | Find the date, look for a newer edition |
| 10 | Second-hand source | Only a checker | An unverified statement | Follow links to the original |
| 11 | Reference that leads nowhere | Only a checker | An invented or broken source | Click every link, resolve every DOI |
Signs 1 to 5 are matters of taste and habit. Signs 7 to 11 are matters of fact, and anyone who doubts your article can show each of them with one link.
“Readers judge by style, so style is what matters”
Readers do react to the visible signs, and a text full of them loses trust before anyone checks a number. So fix the style. Doing so costs minutes, and the six signs above tell you where to look.
But the two repairs are independent. Rewriting a paragraph does not touch the figure inside it, so a wrong number survives every edit of the prose around it. And a polished draft is more dangerous than a clumsy one, because the visible signs were all that prompted anyone to look closer.
Scale matters too. For a study by the European Broadcasting Union and the BBC, published in October 2025, journalists from 22 public broadcasters in 18 countries graded AI assistants’ answers to news questions. They report that “45% of all AI responses were found to have at least one ‘significant’ issue,” and sourcing was the biggest cause, with “31% of all responses having significant issues with sourcing” (EBU and BBC, 2025). Nobody in that study was grading style.
A ten-minute check before you publish
Run it in this order, which follows the cost of the error and not how easy it is to spot.
- List every figure, quotation and named fact in the draft. Our nine articles had between 12 and 71 such items each.
- Open each source. A link that fails or a DOI that does not resolve ends the claim (sign 11).
- Search the page for the quoted words or the number. Not found means not said (sign 7).
- Read around the number. Note what was counted, out of what, and when (sign 8).
- Check the date and look for a newer edition (sign 9).
- Trace second-hand claims to the original and cite that (sign 10).
- Then edit the style for signs 1 to 5, and delete any chat leftovers (sign 6).
- Write the date of the check on the page.
People skip step 1 more than any other, and without the list the rest does not happen. We describe the same work as a stage of production in our guide to fact-checking AI content, and where it fits among the other steps in content writing for SEO.
Frequently asked
How can someone tell if writing is AI?
By style alone, often they cannot. Untrained readers perform near chance, and detectors flag human writing too. Reliable evidence is of another kind. Look for leftovers from the chat window, references that do not exist, and quotations or figures that are not in the cited source.
What does AI writing sound like?
Even in rhythm and slightly promotional in tone. Typical drafts lean on a stock vocabulary, state the importance of ordinary things, attribute views to unnamed experts and close on a hopeful summary. Each of these also appears in human writing, so treat several together as a reason to look closer.
Can AI detectors reliably identify AI writing?
Not well enough to act on a score alone. Seven detectors in one study wrongly flagged more than half of a set of essays by non-native English writers, and OpenAI retired its own classifier, citing its low accuracy. Treat a score as a reason to check the sources.
Does removing the signs of AI writing make a text accurate?
No. Style and accuracy are separate. Editing the wording leaves every figure and quotation exactly as it was, and Wikipedia’s guide warns that fixing only the surface signs “could just make detection harder.” Accuracy comes from opening the sources.
Does Google penalize AI-written content?
Google’s spam policy targets content made at scale with “little to no value to users, no matter how it’s created”, so the method is not the test. Its guidance on generative AI adds a requirement about accuracy: “It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing.”
Sources
14 sources
Methodology: every figure and quotation above was checked against the page it is cited from on November 11, 2026, and each quoted sentence appears on that page. The Wikipedia project pages are quoted from their revisions of November 10, 2026; they change often. Our own counts come from the published records under our nine earlier articles. This article was written with AI assistance and checked the same way.
- 01Wikipedia. “Wikipedia:Signs of AI writing.” Project page, revision of November 10, 2026. https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing.Supports: the six groups of visible signs and their words to watch; the warning against fixing only the signs; the em dash filed under historical indicators; broken links as a strong sign.
- 02Wikipedia. “Wikipedia:Speedy deletion”, criterion G15. Policy page, read November 11, 2026. https://en.wikipedia.org/wiki/Wikipedia:Speedy_deletion#G15.Supports: communication intended for the user and non-existent references as grounds for deleting a page as LLM-generated.
- 03Russell, J., Karpinska, M. and Iyyer, M. “People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text.” 2025. https://arxiv.org/abs/2501.15654.Supports: 300 articles; frequent users’ majority vote wrong on 1 of 300; annotators who do not use AI for writing near chance; vocabulary and formulaic structure as the common giveaways.
Show all 14 sourcesShow fewer
- 04Kobak, D., González-Márquez, R., Horvát, E.-Á. and Lause, J. “Delving into LLM-assisted writing in biomedical publications through excess vocabulary.” 2024, revised 2025. https://arxiv.org/abs/2406.07016.Supports: more than 15 million abstracts; at least 13.5% of 2024 abstracts processed with LLMs.
- 05Geng, M. and Trotta, R. “Human-LLM Coevolution: Evidence from Academic Writing.” 2025. https://arxiv.org/abs/2502.09606.Supports: words identified as overused by ChatGPT dropped in frequency after they were pointed out in early 2024.
- 06Liang, W., Yuksekgonul, M., Mao, Y., Wu, E. and Zou, J. “GPT detectors are biased against non-native English writers.” 2023. https://arxiv.org/abs/2304.02819.Supports: seven detectors, 91 TOEFL essays, average false positive rate 61.22%.
- 07OpenAI. “New AI classifier for indicating AI-written text.” January 31, 2023, with a note of July 20, 2023. https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/.Supports: 26% of AI-written text identified, 9% of human text mislabeled; the classifier withdrawn for low accuracy.
- 08Google Search Central. “Google Search’s guidance on generative AI content on your website.” Last updated October 1, 2026. https://developers.google.com/search/docs/fundamentals/using-gen-ai-content.Supports: generative models predict likely words and do not retrieve facts; AI-generated content must be fact-checked before publishing.
- 09Google Search Central. “Spam policies for Google web search.” Last updated August 28, 2026. https://developers.google.com/search/docs/essentials/spam-policies.Supports: scaled content abuse is defined by lack of value to users, however the content is created.
- 10Wu, K. and colleagues. “An automated framework for assessing how well LLMs cite relevant medical references.” Nature Communications, April 16, 2025. https://www.nature.com/articles/s41467-025-58551-6.Supports: seven models, 800 questions; 50% to 90% of responses not fully supported by their sources; about 30% of statements unsupported for a model with web search.
- 11Walters, W. H. and Wilder, E. I. “Fabrication and errors in the bibliographic citations generated by ChatGPT.” Scientific Reports, September 7, 2023. https://www.nature.com/articles/s41598-023-41032-5.Supports: 55% of GPT-3.5 citations and 18% of GPT-4 citations fabricated; errors in the real ones.
- 12Jergas, H. and Baethge, C. “Quotation accuracy in medical journal articles: a systematic review and meta-analysis.” PeerJ, 2015. https://pmc.ncbi.nlm.nih.gov/articles/PMC4627914/.Supports: 28 studies; a total quotation error rate of 25.4%.
- 13Simkin, M. V. and Roychowdhury, V. P. “Read before you cite!” 2002. https://arxiv.org/abs/cond-mat/0212043.Supports: the estimate that about 20% of citers read the original paper.
- 14European Broadcasting Union and BBC. “News Integrity in AI Assistants: An international PSM study.” October 2025. https://www.ebu.ch/Report/MIS-BBC/NI_AI_2025.pdf.Supports: 45% of AI responses with at least one significant issue; 31% with significant sourcing issues.